Simular
Simular's Sai is a computer-use agent that operates a private cloud desktop or your own machine the way a person does — clicking, typing and navigating real interfaces — to run repetitive workflows across software that exposes no API.
What is Simular?
Simular Inc. calls itself the autonomous computer company, and its flagship product, Sai, is a computer-use agent: software that turns a private cloud virtual machine — or your own device — into a machine that operates itself. Rather than calling APIs, Sai reasons and acts inside real interfaces, clicking and typing through browsers and desktop applications the way a person would, and falling back on APIs, terminals and its own code when that is quicker. Three claims anchor the homepage: it works like a human, it runs on your own remote desktop, and it stays transparent, double-checking with you before any critical action.
The technical stance is deliberately contrarian. Agents built purely on large language models are described as good explorers but poor executors; classical RPA is stable but rigid, confined to predefined actions and fixed APIs. Simular takes a neuro-symbolic route instead, pairing the flexibility of language models with the precision of symbolic code and continual reinforcement learning. One consequence is billed as an economic advantage: repeated runs get faster and cheaper than the ones before them. Another is memory — the virtual machine is a permanent environment where logins, settings and context survive between sessions.
The benchmark claims are published rather than hinted at: 90.1% on WebVoyager for browser use, 72.6% on OSWorld for computer use, and 71.6% on AndroidWorld. The underlying research framework, Agent S, is open source on GitHub and has moved from 20.6% to 69.9% on OSWorld across three generations, against 72% for a human; it took a best-paper award at the ICLR 2025 Agentic AI for Science workshop.
The team comes from DeepMind, Google, Baidu, LLNL and UC Santa Barbara, with more than a hundred papers and ten thousand citations behind it, and both co-founders are ex-DeepMind. The company raised a $21.5 million Series A in December 2025 and claims more than 12,600 sign-ups. Sai runs on macOS, Windows, Linux and virtual machines, in the cloud or on your own hardware, and exposes a documented HTTP API and CLI. Access is currently by invitation.
What it does
- Carry out a task described in plain language on a real computer
- Click, type and navigate desktop applications and browsers exactly as a person would
- Run on a private cloud desktop while your own laptop stays closed
- Replay a workflow it has already learned, on a schedule and unattended
- Ask for confirmation before executing any critical action
- Take instructions remotely through iMessage, SMS or Telegram
- Be driven programmatically through the HTTP API or the sai command-line tool
When to use Simular / When not to
A quick filter to help you decide if Simular is the right fit.
When to use Simular
- Operations and back-office teams whose work spans several tools that do not talk to each other
- Solo builders, founders and freelancers who want unattended automation running on their own machine
- Anyone stuck with software that exposes no public API and cannot be automated through integrations
- Recruiters, sales development reps and account managers repeating the same sourcing and follow-up routine every week
- Developers and QA engineers who want Claude Code, test runs or CI chores supervised on a cloud desktop
When not to use Simular
- Organisations needing a documented GDPR posture: the site never mentions GDPR, offers no DPA and names no EU representative
- Teams that must know in which country their data is hosted, since no hosting location is disclosed anywhere
- Buyers who want an annual contract, as only monthly billing is offered
- Anyone under the age of 18, who is explicitly excluded from the service
- Users who cannot wait for an invitation code, because sign-up is not yet open to everyone
How to use Simular
A typical end-to-end flow, from setup to results.
- Request access: sign-up is invite only, and an invitation code opens the application at sai.simular.ai
- Sign in with a Google account, which is mandatory and grants the agent access to the Google APIs you authorise
- Answer the short 'How did you hear about us?' form that precedes the download
- Choose where the agent runs: a cloud desktop provided by Simular, or your own Windows, macOS or Mac mini machine
- Describe the task in plain English, pick an existing workflow, or simply state the outcome you want
- Watch the agent work and approve the critical steps it flags before it executes them
- Redirect or correct it mid-run in natural language whenever it drifts off course
- Teach it a task once, then set it to rerun on a schedule
- Steer running tasks from your phone through iMessage, SMS or Telegram
- For programmatic use, generate an API key and call the HTTP API or the sai command-line tool
Pros & Cons
Pros
- Reaches software that has no API at all, where integration-based automation simply cannot go
- Covers the whole operating system rather than the browser tab alone
- Runs around the clock on a cloud desktop, so your own machine does not need to stay open
- Costs fall with repetition: the same task gets faster and cheaper on each rerun
- Benchmarks are published and the core agent framework is open source and independently checkable
- Serious research pedigree and funding, with ex-DeepMind founders and a $21.5 million Series A
- A properly documented public API and command-line tool, not just a consumer application
Cons
- Access is by invitation only, so you cannot simply sign up and start
- The USD 20 entry price is an early-access discount on a USD 200 list price, on a plan marked limited supply
- That USD 20 buys prepaid credits rather than unlimited use, so heavy automation costs more
- No GDPR statement, no Data Processing Agreement, no EU representative and no published sub-processor list
- No hosting country or region is disclosed anywhere on the site
- Training on your inputs and outputs is on by default, and opting out means sending an email
- The legal documents are inconsistent: a dead footer link, a JavaScript shell serving a PDF, and readable terms from 2024 that still describe the previous desktop product
Pricing & Plans
A permanent free tier exists, but on the research side of the catalogue rather than on the flagship product: Agent S is open source and Agent S Cloud gives every user a free public computer subject to a queue. Sai itself has no free plan. The cheapest paid entry point is USD 20 per month per computer, presented as an early-access discount on a USD 200 list price and marked limited supply, and that amount is credited entirely as prepaid credits. There is no annual billing. A full refund is available if the subscription is cancelled within twenty-four hours of the initial purchase.
- 10
- 000 reloadable credits
- one developer agent
- macOS and Windows
- workflow editor in developer mode
- Community Gallery sharing
- community support
- a cloud Windows desktop that sleeps when idle
- an always-on Windows or Mac cloud computer
- USD 100 of credits included each month
- native application automation
- additional computers billed separately
- and priority support
- everything in Plus
- unlimited credits
- one free deployment agent
- virtual machine support
- private and team workflow sharing
- full SimuLang support
- full API access
- zero data retention from LLM providers
- everything in Pro
- managed scaling
- all platforms including private cloud
- an advanced workflow editor with custom extensions
- team sharing with audit logs
- custom SimuLang modules
- SOC 2
- HIPAA
- a free public computer for all users subject to a queue
- or a dedicated private computer from USD 50 per month
Data, GDPR & hosting
A consolidated view of how Simular handles your data.
GDPR overview
There is no GDPR implementation to report: the word never appears anywhere on the site, in any page collected. The privacy policy is built entirely on United States state law — Depending on your state of residency, you may have certain rights — and grants access, portability, correction, deletion and an opt-out of targeted advertising, profiling and sale. Rights are exercised by email to support@simular.ai, subject to identity verification, and Simular may decline a request it cannot verify. No EU representative is designated under Article 27, no data protection officer is named, no Data Processing Agreement is offered or mentioned, and no list of sub-processors is published. Users under 18 are excluded outright. European buyers should treat the tool as having no documented GDPR posture.
Who owns the data?
Simular, Inc. collects and controls the data. Beyond the identity and contact details taken from the mandatory Google sign-in, it collects your prompts and the agent's outputs, chat and browsing history, screenshots of the pages and videos you view while the agent runs, device and connection data, and real-time location if you consent to it. Inputs and outputs may be uploaded to Simular's own or to third-party cloud servers. Original voice recordings stay on the device. Google user data is walled off by a dedicated section of the policy: never used for training, never shared for advertising. The 2024 desktop terms state that input and output data are stored locally.
Reuse rights
Simular grants a non-exclusive, non-transferable licence: the software is licensed, not sold, and reverse engineering, redistribution and server use are forbidden. Nothing in the documents transfers ownership of your inputs or outputs away from you, or restricts what you do with the results the agent produces. In the other direction, Simular may use your inputs and outputs to train its models and improve the service unless you opt out by emailing support@simular.ai — a request handled by mail rather than by a switch inside the product. Even after opting out, material flagged for safety review or explicitly reported can still be used. Content is de-identified before training, though flagged material may be re-identified to enforce the terms. Data sent to the named third-party model providers, Anthropic Claude and Google Gemini, is contractually barred from training general-purpose models. Any feedback you volunteer becomes a worldwide, royalty-free licence for Simular.
Data retention & training
Hosting summary
Simular discloses no hosting jurisdiction at all: no country, no region, no data centre provider. The privacy policy states only that data provided directly or collected automatically, inputs and outputs included, may be uploaded to our own or third-party cloud servers, and that everything is held on secure servers behind firewalls. Sai's virtual machine is described as private and isolated, but its location is never given. Data is transmitted to third-party model providers, named as Anthropic Claude and Google Gemini, for the time needed to process a request, then deleted or anonymised under their own commitments. Two things do stay local: original voice recordings are kept on the device and never uploaded, and the 2024 desktop terms state that input and output data are stored on your own computer, with only anonymised statistics transmitted. The website itself sits behind Cloudflare, which says nothing about where the application or its data actually run. Buyers with residency requirements will have to ask directly.
Where Simular works
Country-level availability.
Not available in
Things to keep in mind
Risks and trade-offs to weigh before adopting Simular.
- An agent that operates your desktop needs open sessions and live logins; anything it can click it can also click by mistake, and the terms place the consequences of confirmed actions on you
- Screenshots of the pages and videos you view are collected while the agent runs, which turns ordinary browsing into recorded material
- Your inputs and outputs feed model training by default; the opt-out is an email rather than a switch, and it does not cover content flagged for safety review
- Nothing discloses where the data is hosted, and no GDPR framework, DPA or sub-processor list is published, which is a real obstacle for European or regulated work
- Delegating recurring judgement calls to an agent erodes the practice of making them: the skill you stop exercising is the skill you lose
- Liability is capped at one US dollar and disputes go to individual arbitration in California, so there is little recourse if an automated run causes damage
- The entry price is promotional and credit-based, and an agent that reruns tasks unattended can consume far more than the monthly allowance without anyone watching
Setup & Integrations
Technical difficulty
Low for ordinary use, higher only if you go deep. Simular claims zero setup and no API hooks: you sign in with a Google account, use an invitation code, then describe the task in plain English or pick an existing workflow. Bring-your-own-device means downloading a desktop application for Windows or macOS. The agent asks for confirmation on critical steps, so you supervise rather than configure rules. The technical layer — the developer-mode workflow editor, the SimuLang scripting language, API keys and the command-line tool — exists for those who want it and can be ignored otherwise.
Deployment
Integrations
Behind Simular
Fundraising
Social
Resources
All the official URLs gathered for verification and reference.
Alternatives
Tools that compete with or complement Simular.
Frequently asked questions
What exactly does Simular's Sai do?
Do I need to keep my own computer switched on?
How much does it cost, and is there a free plan?
Is there an API?
Will my data be used to train Simular's models?
Is Simular GDPR compliant?
Can I get a refund?
Who can sign up, and is there an age limit?
Where is my data hosted?
How do I get help?
Should you pick Simular?
Simular is one of the more technically credible entries in the computer-use agent field. The benchmarks are published rather than alluded to, the Agent S framework behind them is open source and independently checkable, the founding team comes from DeepMind, and a $21.5 million Series A led by Felicis gives the company runway. The positioning is unusually clear too: the whole operating system rather than the browser tab, and specifically the software that exposes no API and therefore cannot be reached by integration platforms.
The commercial picture deserves caution. Access is still by invitation. The headline USD 20 per month is an early-access discount on a USD 200 list price, on a plan labelled limited supply, and it buys prepaid credits rather than unlimited use, so the real cost depends on how much the agent actually runs. There is no annual billing at all.
The legal picture is where the gap is widest. Nothing on the site mentions GDPR; there is no Data Processing Agreement, no EU representative, no published sub-processor list and no disclosed hosting country. Training on your inputs and outputs is on by default, and opting out means sending an email. SOC 2 and HIPAA appear only as feature lines on the Enterprise tier. The contractual documents are themselves inconsistent: a dead footer link, a JavaScript shell serving a PDF, and readable terms dated August 2024 that still describe the earlier desktop product.
None of that disqualifies the tool, but it frames the decision. An agent that watches your screen and clicks on your behalf needs live sessions and open logins. For a solo builder automating personal chores that is a reasonable trade; for a regulated European team it is a question the vendor has not yet answered.
- Choosing a selection results in a full page refresh.
- Opens in a new window.